201013546 九、發明說明： 【發明所屬之技術領域】 種根據待測影像中的臉部 一種人臉辨識方法，特別有關於一 角度進行階層式的人臉辨識方法。 【先前技術】 人類臉部辨識系統在近幾年廣泛受到研究學 度重視，亦抑麟計算機裝置㈤如為數位相機、t人電界ΓΙ) 對於數位树的人臉觸上財優異的表現。但是心 的人臉辨識纽，待測影像中的人臉會㈣不201013546 IX. Description of the invention: [Technical field to which the invention pertains] A face recognition method according to a face in an image to be tested, in particular, a hierarchical face recognition method for an angle. [Prior Art] The human face recognition system has been widely valued in research in recent years, and it is also a computer device (5), such as a digital camera, a t-light industry.) The face of the digital tree touches the financial performance. But the face recognition of the heart, the face in the image to be tested will (4) not
致計算機裝置對人臉判斷的錯誤。 用度向V 白知的人臉辨識流程可以區分成訓練與辨識兩個部分 知的人臉辨識的調練過程，係對所有訓練樣本中的臉 進行分類學f。舉例來說，假設臉部角度係幻度作為分_位 的話’那麼0度〜360度則被區分成360個分類區間。計算 曰對所有的雜樣本分獅進行每—個肢辨識訓練。、 除了角度辨識訓練的過程中會產生上述關題外，在進行臉 P角度的觸抵也會歧相應的關。承續上例，因為在訓練 的過程中係糊丨度作為分義區間。所以對於每—個辨識樣本 ，要^做次的臉部角度辨識，從36G次的分類結果:找出 最適田的結果。以M張辨識樣本對N個分類區間進行學習訓練/ 辨識而吕’每—次的角度的訓練/辨識的複雜度係為(M*N)。使得 201013546 這樣=作法會導致下列問題].需要耗費冗長的訓練時間；2.需耗 費的錢體空間。這樣—來，在臉部賴識過程巾就會耗費 許多時間在進行其他不需要的臉部角度麟，並浪 降 體空間與時間。 、 已憶 【發明内容】 鑒於以上的問題，本發_主要目的在於提供—種人、 ❺之階層式訓練方法，利用計算機裝置對複數個钏練樣本進_: 階層中的子影像偵測器之人臉特徵訓練。 為達域目的’本發騎赌之人_叙崎式訓 ^列步驟：__本。執行訓練手段，根據複數個^ 用以將訓練樣本細分至複數個子影像義，並對每 ， 類別相應的子影像偵測器進行複數個人臉特徵之訓練。= 子影像類別執行訓練手段，藉以產生子影像類別 ^ ❿影像_，直至符合細分條件為止。 層之子 從本發明的另一觀點，本發明提出一種 子影像_器對待測影像進行階層式人臉辨識方法^用 機裝置對待測影像進行各子影像偵測器的人臉特徵辨識。十异 =上述目的，本發鴨揭露之—種利用上述訓練方 =偵測器對待測影像進行階層式人臉辨識方法包括=子 載入-待測影像。將同一層中的每一該 下步驟. 測影像進行該人臉辨識手段，用以取得每1二=該待 201013546 .，測影像之-相似度，再從該些子影像_时選料該相似度 取_該子影像制器。重複對所選出的該子影像_器對待測 ^=:r怖㈣制峨峨-層的該些 從本發明的另一觀點，本發明提出—種對待測影像的階層式 =^_物細細输㈣像偵測器 目的’本發糊露之職影像的階層式人臉辨 據触列步驟··取得複數個訓練樣本。執行訓練手段，根 個角度區間用以將訓練樣本細分至複數個子 ::::像_應的子影像細進行複數個人臉特徵之； 的次―:之:二:類:執行訓練手段’藉以產生子影像類別 人層之子衫像類別，直至符合細分條 、 執行人臉辨識手段，根據待測影像用以同=知像。 :影像偵測器。重複對所選出的子影物器==:的 如辨=將ί 繊後一層的子影像_為:。人 本發明將具有相近的臉部角度之辨識樣本 別的分類依據。再將各子影像_遞迴的建㈣肝衫像類 y使得根影像集合與子影像細 者，將待測影像與各個子影像類別中之相應的臉部角㈣2 201013546 歸類至與其相近的子影像類別中，再重複 直至待測影像被歸類到具有符合的臉部角 如此一來，就不需要重複的對每一個角 有關本發明的特徵與實 明如下。 ❹ 【實施方式】 發月ir、為種對待測影像的階層式人臉辨識方法，其利 處理a的電子裳置(例如，數位相機或個人電腦等)進行各子 與辨識兩個部分御=明的人臉辨識流程亦可分為訓練 的影像偵卿⑽Λ 輸物—峨式架構 ( teCt〇r)，母—個影像偵測器根據不同的臉部角度 Ο 比對，使得待測影像被 遞迴的進行比對程序， 度之子影像類別中為止 度進行分類的動作。 作，茲配合圖示作最佳實施例詳細說 ^ A層的子測i'。並且分賴每-個影像侧器 進=臉辨識_練，㈣產生每—個子影像_器的人臉特徵。 17月參考「繁1 {S| ^ ^ 」不，其係分別為本發明人臉辨識之階層 11”.方法之運作流㈣。人臉辨識之階層式訓練方法包括下列 =…取侍複數個訓練樣本(步驟S110)。判斷訓練樣本的例 Z數量衫滿足職(步驟⑽）。若鱗樣本林包括有人臉 :則將樣樣本細分至例外類別(步驟S121)，再完成步驟⑽ 後執行步驟S130。 lJ卜類別的數量超過閥值時，則執行訓練手段（步驟 201013546 S13〇) ’根據複數個角度區間用以將這些·樣本細分至不同的子 影像類別中，並對每一個子影德井 餘化料 別相應的子影像偵測器進行複 數個人祕叙訓練。接著，判斷崎樣本 _別(步驟S140)。若還有子影像⑽^于汾傢類 ^snn 、、j未！過訓練手段時，則重複步 ‘驟S13G直至元成所有子影像類別為止。 請參^度第區=的是场在影像中的與水平線的夹角角度區間。 ❹ ' &」所7^ ’其係為人臉旋轉方向示：S®。對於不同 ^臉鶴^奴對應的纽_。在本發财所指的人臉旋 施_==為域’再以主軸與水平線的夾角。在其他的實 ’二，部角度也可叹主軸與垂直線的夾角。除此之外， 人臉凝轉也可以是人朗轉麵度。請參考「㈣圖」所示，立 係為人臉轉頭方向示意圖。 /、 ^影像偵測器用以侧相應的子影像類別的人臉特徵。人臉 ❹^ ’_用„嶙樣本的整體人臉作為人臉特徵；也可以從訓練樣 本的人臉中選擇至少一眉毛、耳朵、眼睛、鼻子或嘴巴之相對位 :臉特级。細分條件由這些子影像類別之生成階層數量， 亦或疋鱗-相這好影細職量所妓。 成斤有子衫像類別的訓練手段後，即可產生呈 的子影像偵測器。請參考「第3圖」所示，其係為本發日^之 降層式人臉辨識架構示意圖。在「第3圖」最上方的係代表的是 所有丨東樣本的集合’在此將其稱之為根影像集合310。根影像集 201013546 剛⑽㈣子影像類別 中。根據細二=丨=1樣本=分至所屬的子影像類別320 測、、對子#像類別32G相應的子影像偵 類別320再—次^騎母一個生成的子影像 _ ' 生相應的子影像類別32〇〇 在完錢^场__後，g卩可 臉辨識方法。請參層式人 臉辨^ ^ 4 l所不’其縣本發_階層式人 …之運作流程圖。载入待測影像(步驟 辨識手段_卿用以比對同一層中 := Γ=败树，終層^_卿=2 L 一㈣边㈣之子影像偵測器。其中，子影像侧器係選擇 乂 一人臉特徵作為處理相似度之依據。若臉部辨識手段之處理 2不符合臨界條件時，以相似度最高的子影像制輯待測影 純订人臉_手段(步驟⑽）。若臉部_手段之處理結果符 奸顧料，制影像細細耕_巾(倾顯）。重 複對所選出的子影像制器對制影像執行人臉辨識手段，直至 待測影像完成最後-層的這些子影像_器為止(步驟s例。 為了能清楚解說本實施例係以樹的結構來解釋本發明人臉辨 識之階層式繼方法與辨識方法。請參考「第5圖」所示其係為 本實施例之階層示意圖。根影縣合的細岐細係以0 度到360纟。在此以生成三個子影像類別為例，影像集合31〇 201013546 根據三個紐㈣細分成三個子影像酬。三釘影像類別分別 為第一子影像類別510 '第二子影像_ 52〇與第三子影像類別 530 ’將這三個子影像_視為同—層的子影像類別。第—子影像 類別5K)的角度區間設定為〇度到12〇度；第二子影像類別， 的角度區間設定為121度到度；而第三子影像類別53〇的角 度區間設定為241度到360度。 錢’將臉部角度為0度到120度的訓練樣本分類至第一子 影像類別510中’根據這些訓練樣本進行第—子影像類別51〇的 子影像制器之人臉特徵訓練。其巾人臉特徵的訓練方法可以是 j一不限疋為支持向量機(supp〇rt彻恤腿址$谓)、類神經網路 或主軸元件分析方法等。同理，將臉部角度為⑵度到度的 訓練樣本分類至第二子影像類別520 _，並進行第二子影像類別 520的子影像偵測器之人臉特徵訓練。將臉部角度為241度到勤 藝度的訓練樣本分類至第三子影像類別53G中，並進行第三子影像 類別530的子影像偵測器之人臉特徵訓練。 接下來，以第一子影像類別510為例再進行步驟sl3〇，將第 子〜像類別51〇細分出三個子影像類別，其係分別為第四子影 像類別511帛五子影像類別512與第六子影像類別M3。第四子 影像類別511的角度區間設定為〇度到40度；第五子影像類別512 的角度區間設定為41度到80度；而第六子影像類別513的角度 °° 疋為81度到120度。將第四子影像類別511的訓練樣本依 11 201013546 據^臉角度，將其分類至對應的子影像類別中 各自的進行其子影像侧器的人臉特徵爾=像類別在 影像類別執行步㈣3。’直至符合細分條件為止每= 有的訓練樣本會逐—相產生不_子影像 ^ 猶如樹狀結構—樣。 1具組成關係 ❹ 參 r辨人臉賴之階層式崎方法所剌的結果進行人 1制職絲雜集細柯影_別進行相 似度“。同—相子影像制錄據人臉特徵進行待測 =算。特別的是，在進行相似度計算時子影像二 取心的人臉特徵作為相似度處理的依據。舉例來說，奸^ 像制4要進行1G項的人臉特徵的訓練，則會從這 = 取一部份的項蛾全部項规猶·她度計算。 、l ==測影像中的臉部角度係為75:，並將根影像集合 歧=編料㉟物人臉特徵相似 3二子影像類別510中具有根影像集合臉部角度為。 又Ji度的_樣本。因為待測影像中的臉部角度為乃度，所 =ΓΓ類別51G的子影像偵測器之相似度會高於其他子影 像類別。因而選擇第一子影像類別510作為進行下一次辨識的子 景嫌類別。接著，將娜像與第—子影像她的所有特徵進 盯人臉辨齡段。若人麟齡段之處雜料合邮條件時， 則將制影像細分至例外類別中。轉條件的設定係根據各項人 12 201013546 Z欲之差騎決I若是人臉辨識手段之處理絲*符合臨界 Μ時’則進行次—層子影像類別對待測影像的相似度計算。 同理’將待測影像與第一子影像類別训所生成的子影像類 (別為第四子衫像類別51卜第五子影像類別Η2與第六子影 J像類別513)進行相似度的比對。並第四子影像類別Μ卜第五子影 像類:512與第六子影像類別513中選出與待測影像相似度最高/ ❹；Μ像類W在本實施例中，與待測影像相似度最高的係為第 姓/像類別512。接著，將待測影像與第五子影像侧器的所有 二::人臉辨識手段。因為第五子影像類別512係為本實施例 、、層的子影像類別。所以可以判斷待測影像中係具有W度 到8〇度人臉的影像。若是要更為精準職出人臉角度，可以設定 更為多層的子影像類別。 7本發明係重複的選出次—層的子影像類別，並逐—的分層比 參播測办像與&些子影像類別的相似度。所以本發明不需對全部 2子影像_11雜輯_作。轉從所勒辭影像類财 遞迴的進仃分類，就可以避免對不相關的偵測角度進行分類。本 發明的複雜度係為log M *N，與習知分類的複雜度_)有 顧的降极。 本發明將具有減的臉部纽之觸樣本作為柯子影像類 崎，各子影像_遞_建立其賴的子影像類 別，使練雜集合與子鱗_的組顔伽彡成—雜結構。 13 201013546 複遞Causes the computer device to make a mistake in judging the face. The face recognition process with degree to V Baizhi can be divided into the training process of training and recognizing two partial face recognition processes, and the face in all training samples is classified f. For example, suppose the face angle system illusion is divided into _ bits, then 0 degrees to 360 degrees are divided into 360 classification intervals. Calculation 曰 Each limb is trained for each limb identification training. In addition to the above-mentioned problems in the process of angle recognition training, the touch of the face P angle will also be correspondingly closed. Continued from the above example, because in the process of training, the degree of confusion is used as a divisional interval. Therefore, for each identification sample, the face angle recognition should be done twice, and the classification result from 36G times: find the best field result. The learning/recognition of N classification intervals with M identification samples and the training/identification complexity of each angle are (M*N). This makes 201013546 = the practice will lead to the following problems]. It takes a lot of training time; 2. The cost of money. In this way, it will take a lot of time to carry out other unwanted facial angles on the facial process, and to reduce the body space and time. I have recalled the contents of the invention. In view of the above problems, the main purpose of the present invention is to provide a hierarchical training method for human beings and human beings, using a computer device to input a plurality of training samples into the _: sub-image detector in the hierarchy Face feature training. For the purpose of Dayu, 'the person who rides the gambling _Sakisaki training】 Steps: __ Ben. The training means is executed, and the training samples are subdivided into a plurality of sub-image meanings according to a plurality of ^, and the corresponding sub-image detectors of each category are trained for the plurality of personal face features. = The sub-image category performs the training method to generate the sub-image category ^ ❿ image_ until the sub-division condition is met. According to another aspect of the present invention, the present invention provides a method for hierarchical face recognition of a sub-image_device to be measured, and a facial feature recognition of each sub-image detector by a device to be measured. Ten different = the above purpose, the hair duck exposed - using the above training party = detector to measure the image for hierarchical face recognition methods including = sub-loading - image to be tested. Performing the face recognition means for each of the next steps in the same layer to obtain a similarity degree for each image of the image to be acquired, and then selecting the similarity of the images from the sub-images Similarity is taken _ this sub-image controller. Repeating the other sub-images of the selected sub-images to be tested, another aspect of the present invention from the present invention, the present invention proposes a hierarchical pattern of the image to be measured = ^_ Fine-transmission (4) Image detector purpose 'Layer-type face recognition step-steps of the image of the hairline. · A plurality of training samples are obtained. The training method is used, and the root angle interval is used to subdivide the training sample into a plurality of sub-children:::: sub-images like _ should be used to perform complex personal face features; the second--: second: class: executive training means' The sub-image category of the sub-image layer of the sub-image category is generated until the sub-divided strip is executed, and the face recognition means is executed, and the same image is used according to the image to be tested. : Image detector. Repeat the sub-image _ of the selected sub-imager ==: as =. The present invention will have similar classification criteria for identifying samples of similar facial angles. Then, the sub-images are returned to the built-in (four) liver-shirt type y such that the root image set and the sub-image are fine, and the corresponding image angle and the corresponding face angle (4) 2 201013546 in each sub-image category are classified to be similar thereto. In the sub-image category, it is repeated until the image to be tested is classified to have a matching face angle, so that it is not necessary to repeat the features and details of the present invention for each corner as follows. ❹ [Embodiment] The tidal ir, a hierarchical face recognition method for measuring images, is used to process the electronic slap (for example, a digital camera or a personal computer) for each sub-and identification of two parts. The face recognition process of Ming can also be divided into the training image detection (10) 输 input-峨 architecture (teCt〇r), the mother-image detector is compared according to different facial angles, so that the image to be tested is The recursive comparison program, the sub-image category in the degree of classification. For the sake of the preferred embodiment, the sub-test i' of the A layer is described in detail. And depending on each image side device = face recognition _ practice, (4) to generate face features for each sub-image _ device. In July, the reference to "繁1 {S| ^ ^" is not the same as the hierarchy of the face recognition of the invention. 11). The operational flow of the method (4). The hierarchical training method for face recognition includes the following =... Training the sample (step S110). Judging the example Z of the training sample satisfies the job (step (10)). If the scale sample forest includes the human face: the sample sample is subdivided into the exception category (step S121), and then the step is performed after the step (10) is completed. S130. If the number of the lJ class exceeds the threshold, the training means is executed (step 201013546 S13〇) 'Based on the plurality of angle intervals, the samples are subdivided into different sub-image categories, and each sub-picture is The remaining sub-image detectors are used to perform a plurality of personal secret trainings. Next, the sub-images are judged (step S140). If there are sub-images (10)^ in the family class ^snn, j is not! For the training method, repeat step S13G until all the sub-image categories of the element are formed. Please refer to the area = the angle between the field and the horizontal line in the image. ❹ ' &"7^ ' It is the direction of the face rotation: S®. For the different ^ face crane ^ slave corresponding to the new _. In the face of this fortune, the face is rotated _== is the domain' and then the angle between the main axis and the horizontal line. In other real terms, the angle of the main axis can also be sighed by the angle between the main axis and the vertical line. In addition, the face can be turned into a person's face. Please refer to the "(4) diagram" for the direction of the face turning head. /, ^ Image detector is used to face the face features of the corresponding sub-image category. The face ❹^ '_ uses the overall face of the sample as the face feature; you can also select at least one relative position of the eyebrow, ear, eyes, nose or mouth from the face of the training sample: face level. The number of generation levels of these sub-image categories, or the scale-phase of this kind of good-looking and fine-level job. After the training method of the child-like shirts like the category, the sub-image detector can be generated. Please refer to " As shown in Fig. 3, it is a schematic diagram of the falling face type face recognition architecture of the present day. The line at the top of "Fig. 3" represents a collection of all the samples of the ’", which is referred to herein as the root image set 310. Root image set 201013546 Just in the (10) (four) sub-image category. According to the fine second = 丨 = 1 sample = sub-image to the sub-image category 320, the sub-image type 32G corresponding to the sub-image detection category 320, then - the sub-image generated by the rider _ ' The image category 32〇〇 is after the money field __, g卩 face recognition method. Please refer to the face of the face to identify ^ ^ 4 l does not 'the county's hair _ class person ... operation flow chart. Loading the image to be tested (step identification means _qing is used to compare the sub-image detectors in the same layer: = Γ = defeat tree, final layer ^_qing = 2 L one (four) side (four). Selecting a face feature as the basis for processing the similarity. If the face recognition means 2 does not meet the critical condition, the target image with the highest similarity is used to compose the face-to-measure (step (10)). The result of the treatment of the face _ means the traitor, the image is finely ploughed _ towel (dip). Repeat the face recognition method for the selected sub-image controller to the image until the image to be tested completes the last layer These sub-images are exemplified (step s. For the sake of clarity, the present embodiment is explained by the structure of the tree to explain the hierarchical method and identification method of the face recognition of the present invention. Please refer to the figure shown in Fig. 5. This is a hierarchical diagram of the present embodiment. The detailed structure of the root shadow county is 0 degrees to 360 degrees. Here, three sub-image categories are generated as an example. The image collection 31〇201013546 is subdivided into three sub-groups according to three new (four) Image remuneration. The three-nail image category is the first sub-image 510 'Second sub-image _ 52 〇 and third sub-image category 530 ' treat these three sub-images as sub-image categories of the same layer. The angle range of the first sub-image category 5K) is set to 〇 to 12 〇 degree; the second sub-image category, the angle interval is set to 121 degrees to degrees; and the third sub-image category 53 〇 angle interval is set to 241 degrees to 360 degrees. Money 'will face angles from 0 degrees to 120 degrees The training samples are classified into the first sub-image category 510, and the facial feature training of the sub-image controller of the first sub-image category 51 is performed according to the training samples. The training method of the facial features of the towel may be an unlimited疋 is a support vector machine (supp〇rt), a neural network or a spindle component analysis method. Similarly, a training sample with a face angle of (2) degrees to degrees is classified into a second sub-image category. 520 _, and performing face feature training of the sub-image detector of the second sub-image category 520. The training samples with the face angle of 241 degrees to the diligent degree are classified into the third sub-image category 53G, and the first The face of the sub-image detector of the three sub-image category 530 Next, the first sub-image category 510 is taken as an example to perform step sl3 〇, and the first sub-image category 51 〇 is subdivided into three sub-image categories, which are respectively the fourth sub-image category 511 帛 five sub-image categories. 512 and the sixth sub-image category M3. The angle interval of the fourth sub-image category 511 is set to 40 degrees; the angle interval of the fifth sub-image category 512 is set to 41 degrees to 80 degrees; and the sixth sub-image category 513 The angle ° ° 疋 is 81 degrees to 120 degrees. The training samples of the fourth sub-image category 511 are classified according to the face angle of 11 201013546 to the respective sub-image categories of the respective sub-image side devices. Face feature = like category in the image category step (4) 3. 'After meeting the subdivision conditions, each training sample will produce a non-sub-image ^ as if it were a tree structure. 1 composition relationship 参 r 辨 辨 辨 辨 辨 赖 赖 赖 赖 赖 赖 赖 赖 赖 赖 赖 赖 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人 人In particular, the face feature of the sub-image two coring is used as the basis for the similarity processing when the similarity calculation is performed. For example, the rape system 4 is required to perform the training of the face features of the 1G item. , from this = take a part of the item moth all the rules are calculated according to her., l == the angle of the face in the image is 75:, and the root image collection = fabric 35 people The face feature is similar to the 3D sub-image category 510, which has a root image set face angle of . and a Ji-degree _ sample. Because the face angle in the image to be tested is a degree, the sub-image detector of the 51G type The similarity will be higher than other sub-image categories. Therefore, the first sub-image category 510 is selected as the sub-view category for the next recognition. Then, all the features of the na-image and the sub-image are marked into the face. If there is a mixed mailing condition at the age of the aging section, the image will be subdivided into examples. In the category, the setting of the transition condition is based on each person's 12 201013546 Z. The difference is if the face is identified by the method of face recognition. * When the threshold is met, the similarity calculation of the image to be measured is performed. Similarly, the sub-images generated by the image to be tested and the first sub-image category training are similar to the fourth sub-shirt image category 51, the fifth sub-image category Η2 and the sixth sub-image J image category 513. The fourth sub-image category is selected from the fifth sub-image category: 512 and the sixth sub-image category 513 are selected to have the highest degree of similarity with the image to be tested/❹; the image type W is in this embodiment, The highest similarity of the image to be tested is the first name/image category 512. Next, the image to be tested and all the second:: face recognition means of the fifth sub-image side. Because the fifth sub-image category 512 is the implementation For example, the sub-image category of the layer. Therefore, it can be judged that the image to be tested has an image with a W degree to 8 degrees. If the angle of the face is more accurate, a more multi-image sub-image category can be set. 7 The invention repeatedly selects sub-image categories of sub-layers The stratification is more similar to the sub-image categories of the sub-images. Therefore, the present invention does not need to be used for all 2 sub-images. By classifying the classification, it is possible to avoid classifying the irrelevant detection angles. The complexity of the present invention is log M * N, which has a definite polarity with respect to the complexity of the conventional classification _). The sample of the face of the face is used as the image of the scorpion image, and each sub-image _ _ _ establishes the sub-image category of the sub-image, so that the group of the squad and the sub-scale _ _ _ _ _ _ _ _ _ _ _ _ _
，直至待測影像被歸類到具有符合的臉部角度 之·^像_中為止。如此—來除了可以利用較少的運算量以及 =憶體容#就可崎算出待測影射是否具有臉部區域。 定本=本她前述之較佳實施例揭露如上，然其並非用以限 内二可作Γ蝴相像技藝者’在不脫離本發明之精神和範圍 才__，㈣本购之專護範圍綱 本說軋所附之中請專利範圍所界定者為準。 【圖式簡單說明】Until the image to be tested is classified into the image with the matching face angle. In this way, in addition to using less computational complexity and = memorizing the body #, it is possible to calculate whether the image to be measured has a face region. The preferred embodiment of the present invention is as disclosed above, but it is not intended to be used by those skilled in the art, and the scope of the invention is not limited to the spirit and scope of the present invention. In the context of the patent, the scope defined in the patent is subject to change. [Simple description of the map]
f 1圖係分顺本發明人臉觸之階層式繼方法之運 程圖。 ；IL 第2a圖係為人臉旋轉方向示意圖。 第2b ϋ係為人臉轉頭方向示意圖。 第圖心、為本發明之階層式人臉辨識架構示意圖。 第4圖係為本發明的階層式人臉辨識方法之運作流程圖。 弟5圖係為本實施例之階層示意圖。 【主要元件符號說明】 310 根影像集合 320 子影像類別 510 第—子影像類別 14 201013546 520 第二子影像類別 530 第三子影像類別 511 第四子影像類別 512 第五子影像類別 513 第六子影像類別The f 1 map is a diagram of the hierarchical method of the inventor's face. ;IL Figure 2a is a schematic diagram of the direction of rotation of the face. The 2b ϋ is a schematic diagram of the direction of the face turning. The figure is a schematic diagram of the hierarchical face recognition architecture of the present invention. Figure 4 is a flow chart showing the operation of the hierarchical face recognition method of the present invention. The figure 5 is a hierarchical diagram of the present embodiment. [Description of main component symbols] 310 image sets 320 Sub-image category 510 First-sub-image category 14 201013546 520 Second sub-image category 530 Third sub-image category 511 Fourth sub-image category 512 Fifth sub-image category 513 Sixth sub-image Image category
1515